SDSC-YOLOv11-seg: a lightweight river oil spill segmentation method

Abstract

Real-time river oil spill segmentation is hindered by extreme target morphological variations, blurred oil-water boundaries, and the difficulty of deploying heavy models on edge platforms. To address these challenges, this paper proposes an efficient lightweight instance segmentation model named SDSC-YOLOv11-seg. The architecture systematically integrates a switchable atrous convolution backbone for adaptive receptive field allocation, a cross-layer compressed feature fusion neck to slash complexity, a dynamic lightweight segmentation head to accelerate decoding, and a spatial-channel synergistic attention module paired with the ShapeIoU metric for precise boundary localization. Experiments on a self-built dataset of 2432 river oil spill images show that SDSC-YOLOv11-seg achieves a detection mAP50 of 94.7% and a mask mAP50 of 93.9%, using only 2.054 × 106 parameters and 6.9 GFLOPs of computation. Crucially, zero-shot generalization tests in cluttered rocky riverbank scenes confirm its ability to capture micro-slicks while avoiding catastrophic over-segmentation. Furthermore, benchmark deployment on an NVIDIA GTX 1650 Ti mobile GPU achieves a real-time throughput of 57.2 frames per second, outperforming the native lightweight baseline. This method achieves an optimal Pareto balance between accuracy and efficiency, offering a robust engineering solution for airborne drone monitoring.